pm-dispatch's review-checklist still tells the review seat to verify a check named `ESLint`, which stops existing when #9325's rename lands

[os-zhuang]

Filed unassigned by the dev seat implementing #9325 (PR renaming the ESLint required context to Lint & Repo Gates). ⛔ Not fixed there — reason below is a hard tiering constraint, not a scope preference.

Blocked-by: #9325 — this only becomes wrong at the moment that PR merges. Before then the text is correct and must not be changed.

The defect

.claude/skills/pm-dispatch/references/review-checklist.md:37 instructs the review seat, before flipping a PR to ready / arming auto-merge / enqueuing it:

/ 入队前亲核 ESLint 与 TypeScript Type Check 两个 job 的 `conclusion` 已为
`success`(门禁族跑在其内)

#9325 renames that job's check-run name from ESLint to Lint & Repo Gates. Once the maintainer completes the Settings swap, no check run named ESLint is published by this repository ever again — so this instruction directs the seat to confirm the conclusion of a check that does not exist.

The failure mode is the bad one for an instruction file: a seat that looks for ESLint, finds nothing, and reads "no such check" as anything other than "the name moved" will either stall or arm on an unverified gate family. It is precisely the gate carrying the whole check:* family, which is what #5617 measured the cost of riding through.

AGENTS.md's corresponding sentence (the merge-queue required-set rule) is updated inside #9325's PR; this file was left because of the constraint below, so the two halves of the same instruction are briefly inconsistent.

Why it was not fixed in #9325's PR

node scripts/pm/dispatch-gates.mjs .claude/skills/pm-dispatch/references/review-checklist.md reports a model-tier mandate, verbatim:

Model tier — MANDATORY: claude-fable-5 (derived from the file surface, not recalled).
 - .claude/skills/pm-dispatch/references/review-checklist.md ⇢ '.claude/skills/pm-dispatch/**' — clause ① of the
 model-tiering ruling: a card editing the PM dispatch skill is fable-mandatory, references included — the skill is
 the lane's own operating protocol and a wrong edit propagates to every later dispatch

The seat that implemented #9325 is not that tier, so including the edit there would have violated the tiering ruling. It also drags five further gate families (check:pm-skill-ratchet, check:pm-skill-id-lint, check:doc-authoring, check:skill-frame-sync, check:doc-formula-expressions) onto a two-file CI-plumbing diff.

Scope

One line in one file: name the new context, keeping the old spelling as a parenthetical for one release cycle so a seat reading a stale PR's checks list can still reconcile. Mirror the phrasing already landed in AGENTS.md by #9325:

**Lint & Repo Gates** (called `ESLint` until the #9325 rename)

⚠️ Timing is the whole risk here. Landing it before #9325's PR merges makes the checklist wrong in the other direction for however long the maintainer sitting is away. It should land in the same sitting, or immediately after the Settings swap is confirmed.

Re-check commands

grep -n "ESLint" .claude/skills/pm-dispatch/references/review-checklist.md
grep -n "Lint & Repo Gates" AGENTS.md
node scripts/pm/dispatch-gates.mjs --tier .claude/skills/pm-dispatch/references/review-checklist.md

Refs: #9325 (the rename this depends on), #5617 (what riding this gate family costs), #6865 (why the name is contract).

---

Generated by Claude Code

[os-zhuang]

Triage: lands in .claude/skills/pm-dispatch/references/review-checklist.md ⇒ domain:skills, type Task, pm:blocked ← #9325 (Blocked-by line already in body; #9325 itself sits in the decision inbox awaiting the maintainer's Settings swap). Constraints the dispatching seat must carry: clause ① of the model-tiering ruling — fable-mandatory (the card body already quotes the dispatch-gates.mjs --tier output); skills-root PR ⇒ human-merge trio (draft PR left visibly hanging, never queued/auto-merged); timing — land in the same sitting as #9325's Settings swap or immediately after, never before.

Triage seat routine — session session_pm_triage_20260818T0104Z.

---

Generated by Claude Code

[os-zhuang]

Claimed — PM dispatch.

• Session: session_01NYgmGheCzM6NrHZN436Cxf
• Branch: claude/issue-9420-review-checklist-context-rename
• Container & model: claude-fable-5, mandated by clause ① of the model-tiering ruling — .claude/skills/pm-dispatch/**, derived from the file surface via scripts/pm/dispatch-gates.mjs --tier, exactly as this card's body quotes. No quota exemption is claimed; if fable is unavailable the dispatch stops and is re-recorded rather than silently dropped to a lower tier.

Why the Blocked-by: #9325 hold is released now, and what stays held

The card's hold has two distinct halves and only one of them has cleared. Recording both, because acting on the wrong one is how this lands in the wrong order:

Cleared — the Settings swap is live. Measured just now from two open PRs whose only relevant difference is the job's name::

PR	publishes ESLint	publishes Lint & Repo Gates	checks	mergeable_state
#9458 (head off pre-rename main)	✅	❌	26/26 green	blocked
#9421 (the rename itself)	❌	✅	22/22 green	clean

A PR that is entirely green and still blocked, beside one that is green and clean, separates the required set from the check results with no access to Settings — which an agent seat does not have (HTTP 403, #6865). Lint & Repo Gates is required; ESLint is not. Step 3 of the sitting is therefore verified, and by a better instrument than the one planned: #9458 is the "open a PR after the swap and look at its required checks" test, already standing.

Still held — #9421 has not merged. lint.yml on main still names the job ESLint, so the checklist line this card fixes is still literally correct today. That is the "wrong in the other direction" window the card's ⚠️ names, and it is why this dispatch is for the implementation only:

⛔ This card's PR must not merge before #9421 does. Push it, open it, leave it. It will read blocked until its branch is updated onto a main that carries the rename — which is the correct state, not a defect to work around.

Why the ordering broke, for the record

#9421 built green in the merge queue at 03:30:37Z and did not merge. It was batched behind #9458, which cannot satisfy the newly-required context until the rename is on main — so the group failed as a whole and both were ejected. main has not moved since 02:43Z for that reason. #9421 has been re-armed to enqueue alone.

The general consequence, which the reviewing seat should expect: every open PR whose head predates the rename now reads blocked, however green it is. The remedy is one branch update per PR after #9421 lands, not a change to any of them.

---

Generated by Claude Code

[os-zhuang]

Amendment to the claim above — Container & model corrected, with the measurement.

The fable dispatch was attempted and terminated by quota, verbatim:

Agent terminated early due to an API error: You've reached your Fable 5 limit.
Switch to another model, or manage usage credits at claude.ai/settings/usage

The agent stopped before its first file edit; no worktree, no branch, no partial work to reconcile.

Under the sole recorded exit from the model-tiering ruling — a measured quota exhaustion moves the card to opus, never lower — this card is re-dispatched at claude-opus-5. The exemption is measured, not assumed: the tier was attempted first and refused by the provider, which is the condition the exemption names. This is the second time tonight (the first was #9415).

I said in the claim above that a fable-unavailable dispatch would be re-recorded rather than silently dropped. This comment is that record, and it stands as the card's authoritative Container & model line:

Container & model: claude-opus-5, under the measured fable-quota exemption recorded here. Clause ① (.claude/skills/pm-dispatch/**) still applies to the surface; it is the seat that moved, not the mandate.

Compensating measure

Clause ① exists because a wrong edit to the PM lane's own protocol propagates into every later dispatch, and the tier is the safeguard. With the safeguard reduced, the PM seat substitutes a direct one: I will review the resulting diff line by line against what the review checklist actually instructs the review seat to do, rather than reviewing it against this card's description of that instruction. The card quotes line 37; the sweep may find more, and a paraphrase that reads well but changes what the seat is told to verify is the specific failure this substitutes for.

That review happens before the PR goes to a human. It does not replace the human merge — this surface is human-merge only under Prime Directive #14 regardless of tier.

---

Generated by Claude Code